Conflicting Biases at the Edge of Stability: Norm versus Sharpness Regularization